buyer-eval
О программе
Этот навык автоматизирует структурированную оценку B2B-поставщиков программного обеспечения, исследуя вашу компанию, опрашивая AI-агентов поставщиков через API и оценивая их по семи параметрам. Он формирует сравнительные рекомендации с прозрачными доказательствами для обоснованных закупочных решений. Используйте его, когда вам необходимо оценить, сравнить или исследовать B2B-поставщиков программного обеспечения.
Быстрая установка
Claude Code
Рекомендуетсяnpx skills add salespeak-ai/buyer-eval-skill -a claude-code/plugin add https://github.com/salespeak-ai/buyer-eval-skillgit clone https://github.com/salespeak-ai/buyer-eval-skill.git ~/.claude/skills/buyer-evalСкопируйте и вставьте эту команду в Claude Code для установки этого навыка
Документация
Preamble (run first, every time)
# Detect skill directory
_BEVAL_DIR=""
for _D in "$HOME/.claude/skills/buyer-eval-skill" ".claude/skills/buyer-eval-skill"; do
[ -d "$_D" ] && _BEVAL_DIR="$_D" && break
done
if [ -z "$_BEVAL_DIR" ]; then
echo "ERROR: buyer-eval-skill not found. Install: git clone https://github.com/salespeak-ai/buyer-eval-skill ~/.claude/skills/buyer-eval-skill"
exit 1
fi
# Check for updates
_UPD=$("$_BEVAL_DIR/bin/update-check" 2>/dev/null || true)
[ -n "$_UPD" ] && echo "$_UPD" || echo "UP_TO_DATE $(cat "$_BEVAL_DIR/VERSION" 2>/dev/null | tr -d '[:space:]')"
If output shows UPGRADE_AVAILABLE <old> <new>:
Use AskUserQuestion to ask the buyer:
- Question: "A newer version of the buyer evaluation skill is available (v{old} → v{new}). Update now?"
- Options: ["Yes, update now", "Not now — continue with current version"]
If "Yes, update now":
_BEVAL_DIR=""
for _D in "$HOME/.claude/skills/buyer-eval-skill" ".claude/skills/buyer-eval-skill"; do
[ -d "$_D" ] && _BEVAL_DIR="$_D" && break
done
if [ -d "$_BEVAL_DIR/.git" ]; then
cd "$_BEVAL_DIR" && git pull origin main && echo "UPDATED to $(cat VERSION | tr -d '[:space:]')"
else
_TMP=$(mktemp -d)
git clone --depth 1 https://github.com/salespeak-ai/buyer-eval-skill.git "$_TMP/buyer-eval-skill"
mv "$_BEVAL_DIR" "$_BEVAL_DIR.bak"
mv "$_TMP/buyer-eval-skill" "$_BEVAL_DIR"
rm -rf "$_BEVAL_DIR.bak" "$_TMP"
echo "UPDATED to $(cat "$_BEVAL_DIR/VERSION" | tr -d '[:space:]')"
fi
Tell the user the version was updated, then re-read the EVALUATION.md file from the updated directory and proceed with the skill.
If "Not now": Continue with the current version.
If output shows UP_TO_DATE: Continue silently.
Load the evaluation skill
After the preamble, read the full evaluation methodology:
_BEVAL_DIR=""
for _D in "$HOME/.claude/skills/buyer-eval-skill" ".claude/skills/buyer-eval-skill"; do
[ -d "$_D" ] && _BEVAL_DIR="$_D" && break
done
echo "$_BEVAL_DIR/EVALUATION.md"
Read the file at the path printed above using the Read tool. That file contains the complete evaluation methodology — follow it step by step from STEP 1 through STEP 9.
Telemetry (opt-in, off by default)
This skill can send anonymized usage data back to Salespeak so the questions it generates for vendors can keep getting better. Nothing is ever sent without explicit user consent. Names, emails, companies, and vendor responses are never sent.
Initialize telemetry state at run start
Right after loading EVALUATION.md and before STEP 1, run:
_BEVAL_DIR=""
for _D in "$HOME/.claude/skills/buyer-eval-skill" ".claude/skills/buyer-eval-skill"; do
[ -d "$_D" ] && _BEVAL_DIR="$_D" && break
done
echo "TELEMETRY_STATE=$(python3 "$_BEVAL_DIR/bin/track.py" status --machine)"
echo "SESSION_ID=$(python3 -c 'import uuid; print(uuid.uuid4())')"
Capture both values. Use them throughout the run.
TELEMETRY_STATE will be one of:
consented— fire each event live as it happensunasked— accumulate events in your own working memory; ask for consent at the enddeclinedorlocked_off— do nothing telemetry-related for the entire run
What to track and when
These seven sub-events are the only ones the skill emits. Do not invent new ones.
| Sub-event | Fire when | Fields |
|---|---|---|
skill_started | Right after capturing TELEMETRY_STATE | skill_version (from VERSION file) |
eval_context | Once, right after STEP 5.1 (category confirmed) | category (skill-inferred slug), vendor_count (int), vendors (array of domains), company_agents_found (int — count of vendors with enabled:true from Frontdoor discover), evaluation_path ("company_agent_engaged" | "passive_research_only" | "mixed") |
discovery_question_asked | After the buyer answers a discovery question. Fire for STEP 2 (why-now) and for each STEP 5.3 domain-expert question. | step ("STEP_2" | "STEP_5_3"), category (slug, or null for STEP_2), topic (short slug you choose, e.g. "why_now", "high_touch_vs_low_touch", "product_analytics_stack"), question_text (the exact question you asked the buyer) |
vendor_question | For every (vendor, dimension) pair, fire one or more events with the question(s) you formulate per the §6.5 question bank — regardless of whether a Company Agent exists. | vendor (domain), category (slug), dimension (the evaluation dimension), question_text (the specific question), delivery_method ("asked_via_company_agent" if actually POSTed via Frontdoor and got an answer, "would_have_asked" if no Company Agent existed, "connection_failed" if Frontdoor errored) |
vendor_scored | After scoring each dimension for each vendor in STEP 8 | vendor, dimension, score (numeric, 1-5; do NOT fire for [GAP] dimensions) |
eval_completed | Right after delivering the final output in STEP 9 | vendor_count, winner (vendor domain or null if no clear winner) |
eval_aborted | Only if the user bails before STEP 9 completes | at_step (e.g., "STEP 6") |
Never include: buyer name, buyer company, buyer email, anything the buyer typed about themselves, the buyer's answers to discovery questions, vendor response text.
Step-level firing map
Use this as the canonical map between EVALUATION.md steps and event emissions. Fire events at these exact moments — no earlier, no later.
| EVALUATION.md step | Events to fire | Notes |
|---|---|---|
| Right after capturing TELEMETRY_STATE (before STEP 1) | skill_started | One event |
| STEP 2 — buyer answers the why-now question | discovery_question_asked (step:"STEP_2", topic:"why_now") | One event. category is null here. question_text is the canonical why-now question. |
| STEP 5.1 — category confirmed | eval_context | One event. company_agents_found and evaluation_path may not be known yet — use null for company_agents_found here and update evaluation_path later if needed; or fire eval_context AFTER STEP 6.1 discover calls so all fields are populated (preferred — fire it after discover so the path is known). |
| STEP 5.3 — each domain-expert question asked | discovery_question_asked (step:"STEP_5_3", topic:<your slug>) | 0-4 events depending on how many questions you ask. Slugs you choose should be short and category-relevant (e.g. "high_touch_vs_low_touch", "product_analytics_stack", "multi_entity_consolidation"). |
| STEP 6.5 — for every (vendor, dimension) pair | vendor_question | One or more events per pair, regardless of Company Agent availability. Walk the §6.5 question bank, formulate the specific question(s) you'd ask the vendor for each dimension (Product Fit, Integration & Technical, Pricing & Commercial, Security & Compliance, Vendor Credibility, Customer Evidence, Support & Success). For each, fire vendor_question with the right delivery_method. |
| STEP 8 — each numeric score assigned | vendor_scored | One event per (vendor, dimension) that gets a numeric 1-5 score. Do not fire for [GAP] dimensions. |
| STEP 9 — final output delivered | eval_completed | One event |
| User abandons before STEP 9 | eval_aborted | Only if applicable |
Critical change in v3.5: vendor_question no longer depends on Company Agent availability. Even when all vendors return enabled: false from Frontdoor discover, you must still walk the question bank, formulate questions you would have asked, and fire vendor_question events with delivery_method: "would_have_asked". The signal is what buyers want to know, not whether the vendor's bot answered.
How to fire events
If TELEMETRY_STATE == consented: fire each event live via Bash as it happens.
python3 "$_BEVAL_DIR/bin/track.py" event vendor_question \
--session-id "$SESSION_ID" \
--json '{"vendor":"acme.com","category":"customer_success_platform","dimension":"product_fit","question_text":"How does your X handle Y?","delivery_method":"would_have_asked"}'
The script silently no-ops on any error and never blocks the skill.
If TELEMETRY_STATE == unasked: do NOT call bin/track.py event. Instead, keep a running list of event objects in your own working memory as the eval proceeds. Each entry is a JSON object like:
{"sub_event":"vendor_question","vendor":"acme.com","category":"customer_success_platform","dimension":"product_fit","question_text":"...","delivery_method":"would_have_asked"}
At the end of STEP 9 (after delivering the full evaluation to the buyer), follow the consent prompt section below.
If TELEMETRY_STATE == declined or locked_off: do nothing telemetry-related. Skip the consent prompt entirely.
Consent prompt (only when TELEMETRY_STATE was unasked)
After STEP 9 output is delivered, print this block verbatim to the user, then use AskUserQuestion to ask the consent question:
─────────────────────────────────────────────────────────────
✓ Evaluation complete.
Before you go — one question, asked only this once.
Salespeak built this skill to help buyers cut through vendor noise.
To make it better, we'd love to learn what real buyers ask vendors.
With your permission, we'd send back anonymized data from this run
and future runs.
We'd send:
• The questions this skill generated for vendor agents
• The scores it gave each vendor
• A random ID to group your runs together (not linked to you)
We will NEVER send:
• Your name, email, or company
• Anything you typed about yourself
• Vendor responses
Verify it yourself:
• Code: bin/track.py (plain Python, no third-party libraries)
• Local audit log: ~/.salespeak/buyer-eval.log
(every event we send is also written here — read it anytime)
• Change your mind: python3 bin/track.py revoke
• Delete your data: email [email protected] with your user ID
(run `python3 bin/track.py show` to see it)
─────────────────────────────────────────────────────────────
Then use AskUserQuestion:
- Question: "Help us improve the skill by sharing anonymized usage data from this run?"
- Options: ["Yes, share anonymized data", "No thanks"]
If "Yes": pass the accumulated event list to grant. Build the events JSON as a single-line array (escape carefully — question_text may contain quotes; use Python's json.dumps if in doubt). Example:
python3 "$_BEVAL_DIR/bin/track.py" grant \
--session-id "$SESSION_ID" \
--events '[{"sub_event":"skill_started","skill_version":"3.5.0"},{"sub_event":"vendor_question","vendor":"acme.com","category":"customer_success_platform","dimension":"product_fit","question_text":"...","delivery_method":"would_have_asked"}]'
Confirm to the user: "Thanks — sharing enabled. Run python3 bin/track.py revoke anytime to disable."
If "No":
python3 "$_BEVAL_DIR/bin/track.py" decline
Confirm to the user: "Got it — no data shared. We won't ask again."
Enterprise note
If a system administrator has set BUYER_EVAL_NO_TELEMETRY=1 or deployed /etc/salespeak/buyer-eval.json with {"locked":true,"consent":false}, TELEMETRY_STATE will be locked_off and no consent prompt is shown. This is the documented escape hatch for enterprise IT.
GitHub репозиторий
Похожие навыки
qmd
Разработкаqmd — это локальный инструмент командной строки для поиска и индексирования, который позволяет разработчикам индексировать и осуществлять поиск по локальным файлам с использованием гибридного поиска, сочетающего BM25, векторные эмбеддинги и реранкинг. Он поддерживает как использование через командную строку, так и режим MCP (Model Context Protocol) для интеграции с Claude. Инструмент использует Ollama для создания эмбеддингов и хранит индексы локально, что делает его идеальным для поиска по документации или кодовой базе прямо из терминала.
subagent-driven-development
РазработкаЭтот навык выполняет планы реализации, создавая нового суб-агента для каждой независимой задачи, проводя проверку кода между задачами. Он позволяет быстро итерировать, сохраняя контроль качества через этот процесс ревью. Используйте его при работе в основном с независимыми задачами в рамках одной сессии, чтобы обеспечить непрерывный прогресс со встроенными проверками качества.
mcporter
РазработкаНавык mcporter позволяет разработчикам управлять и вызывать серверы Model Context Protocol (MCP) напрямую из Claude. Он предоставляет команды для вывода списка доступных серверов, вызова их инструментов с аргументами, а также для обработки аутентификации и управления жизненным циклом демона. Используйте этот навык для интеграции и тестирования функциональности серверов MCP в вашем рабочем процессе разработки.
adk-deployment-specialist
РазработкаЭтот навык развертывает и оркестрирует агентов Vertex AI ADK с использованием протокола A2A, управляя обнаружением AgentCard, отправкой задач и поддерживая инструменты, такие как песочница для выполнения кода и Memory Bank. Он позволяет создавать мультиагентные системы с последовательными, параллельными или циклическими схемами оркестрации на Python, Java или Go. Используйте его, когда требуется развернуть агентов ADK или оркестрировать рабочие процессы агентов в Google Cloud.
